Toward Scalable and Unified Example-Based Explanation and Outlier Detection

Area of Science:

• Artificial Intelligence
• Machine Learning
• Explainable AI (XAI)

Background:

• High-stakes AI decisions require transparent explanations and outlier detection.
• Current methods often treat explainability and outlier detection separately.

Purpose of the Study:

• To unify explainability and outlier detection in neural networks.
• To propose and evaluate prototype-based student networks for example-based explanations and outlier identification.

Main Methods:

• Developed novel iterative prototype replacement algorithm for sampling training data.
• Utilized prototype similarity scores for outlier detection.
• Compared prototype-based networks against baseline models.

Main Results:

• Prototype-based networks provide meaningful example-based explanations.
• Achieved promising outlier detection results.
• Maintained classification accuracy without compromise.

Conclusions:

• Prototype-based networks offer a unified approach to explainability and outlier detection.
• These networks enhance trust and verification in AI systems.
• The proposed method demonstrates effectiveness beyond standard similarity kernels.